CN102313897A - Radioactive spectrum identification method - Google Patents

Abstract

The invention discloses a radioactive spectrum identification method comprising the following steps of: carrying out filtration, dimension reduction and classified treatment on measured energy spectrum in radioactive measurement, extracting feature vectors of energy spectrums as samples, and obtaining the GMMs (Gaussian Mixture Models) of all kinds of feature vectors through training; solving the feature vectors of energy spectrums to be identified and solving the class conditional probability density of the feature vectors according to the GMMs; and then carrying out decision classification to finish energy spectrum identification. The radioactive spectrum identification method has high energy spectrum identification accuracy rate and is an effective method for radioactive spectrum identification.

Description

A kind of radioactive energy spectrum recognition methods

Technical field

The present invention relates to a kind of radioactive energy spectrum recognition methods.

Background technology

In carrying out radioactive spectral measurement, a large amount of random data that usually will collect according to detector are analyzed these random signals and by what radioactive source are produced.General radioactive energy spectrum analytic method is that the power spectrum surveyed is being carried out after smothing filtering, deduction background, peak-seeking, calculating net peak area etc. analyze; Result according to energy, peak shape and efficiency calibration; Identify radioactive source by the pairing energy of peak position and contain which radioactive nuclide, and then calculate the activity of these radioactive nuclides by net peak area.In some application scenario, because nuclear information is protected, need not obtain the precise results of nucleic activity, which kind of nuclear material only needs the identification object of surveying be.For this reason, once there was the researcher to adopt the training sample of spectral information, and these nuclear materials discerned with the neural network after the training as neural network.These methods existence are difficult to train, perhaps are difficult for restraining, perhaps converge to shortcomings such as local minimum; Cause the identification of power spectrum can not carry out or the recognition accuracy reduction; Event need be found out and a kind ofly can quick training can avoid again restraining or the method for local convergence, the identification of power spectrum can be carried out smoothly and can guarantee higher recognition accuracy.

Summary of the invention

The objective of the invention is to disclose a kind of radioactive energy spectrum recognition methods.This method has overcome the deficiency of existing power spectrum recognition methods, has characteristics such as training speed is fast, convergence is accurate, recognition accuracy height, is a kind of effective ways that carry out the radiological performance spectrum discrimination.

The present invention realizes that through following technical scheme concrete steps of the present invention are following:

1. the power spectrum that has recorded in the radioactivity survey is trained, presses following A～E step:

A classifies to a plurality of power spectrums that recorded by the radioactive source classification as required,

B adopts wavelet method or polynomial expression method that power spectrum is carried out smothing filtering,

C asks for the GMM model (Gaussian mixture model) of power spectrum, and with the weights of each Gaussian function in the GMM model data after as dimensionality reduction,

The gamma-spectrometric data of D after with dimensionality reduction carries out WAVELET PACKET DECOMPOSITION, and the energy of each sub-band signal carried out normalization handle, and extracts normalization data as proper vector.

E as sample, adopts expectation maximization method (Expectation Maximization is abbreviated as EM) that these samples are carried out interative computation proper vector, obtains the GMM model of all kinds of proper vectors;

2. with the power spectrum to be identified classification of making a strategic decision, by following A～E step:

A adopts wavelet method or polynomial expression method that power spectrum to be identified is carried out smothing filtering,

B asks for the GMM model (Gaussian mixture model) of power spectrum to be identified, and with the weights of each Gaussian function in the GMM model data after as dimensionality reduction,

The to be identified gamma-spectrometric data of C after with dimensionality reduction carries out WAVELET PACKET DECOMPOSITION, and the energy of each sub-band signal carried out normalization handle, and extracts normalization data as proper vector,

D as the multidimensional random number, and calculates the class conditional probability density that it belongs to all kinds of GMM models with the proper vector of power spectrum to be identified,

E presses the Bayesian decision classification at last.

The invention has the beneficial effects as follows:

The present invention adopts wavelet method or polynomial expression method that power spectrum is carried out smothing filtering earlier in the training stage, to eliminate undesired signal; Then power spectrum is done dimension-reduction treatment, still have most information of original ability spectrum signal with the statistical property that guarantees power spectrum at lower dimensional space with the GMM model; Then, extract the normalization proper vector and, guaranteed to adopt low dimension sample to represent the uniqueness of ability spectrum signature, also guaranteed the independence of sample and Measuring Time simultaneously as sample at the multiband signal space; At last, adopt the expectation maximization method that these samples are carried out interative computation, obtain the GMM model of all kinds of proper vectors, so not only guaranteed that accurate convergence but also Gaussian function number that can be through the reasonable GMM of adjustment model are to satisfy the needs of discerning under the various different occasions.In addition, the present invention adopts the Bayesian decision classification when identification, guaranteed the optimum of recognition accuracy under statistical significance.In a word, the present invention a kind ofly can quick training can avoid again restraining or the method for local convergence, the identification of power spectrum can be carried out smoothly and can guarantee higher recognition accuracy.

Description of drawings

Fig. 1 is the process flow diagram of the inventive method;

Embodiment

For the purpose, technical scheme and the advantage that make invention is clearer, below with reference to the accompanying drawing embodiment that develops simultaneously, the present invention is done further explain.

Because the nuclear information protection need not obtain nucleic activity precise results, which kind of nuclear material only needs the identification object of surveying is, to this situation, the invention provides a kind of radioactive energy spectrum recognition methods.Fig. 1 has shown the flow process of recognition methods according to the invention.

Flow process of the present invention is as shown in Figure 1, specifically comprise following training step 1. with identification step 2.:

1. the power spectrum that has recorded in the radioactivity survey is trained, presses following A～E step:

A presses the radioactive source classification as required---as, the kind of standard source or sample and batch etc.---a plurality of power spectrums to having recorded are classified, and power spectrum be measured under apart from, different measuring time and other different external conditions at different measuring;

B adopts wavelet method or polynomial expression method that power spectrum is carried out smothing filtering;

C asks for the GMM model (Gaussian mixture model) of power spectrum, the linearity that promptly is expressed as a plurality of gauss of distribution function with:

$P(x,\mathrm{\θ})=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{a}_i{p}_i(x;{\mathrm{\θ}}_i)---\left(1\right)$

M is the number of Gaussian distribution density function in the formula (1), should count N etc. by shape, smooth degree, the location, road of power spectrum and decide; a ₁..., a _MBe the weight of each Gaussian distribution density function, i.e. each gauss of distribution function shared proportion in probability density function, and

a _i>=0, (i=1 ..., M), decide by gamma-spectrometric data and function number M; p _i(x) be i Gaussian distribution density function, its average μ _iNumber M and location, road number by the Gaussian distribution density function are confirmed; p _i(x) variance is σ _i ²θ _iBe unknown parameter μ _iAnd σ _i ²Vector representation, promptly

Density function p _i(x, θ _i) form following:

$p_i(x,{\mathrm{\θ}}_i)=\frac{1}{{\left(2\mathrm{\π}\right)}^{1/2}{\mathrm{\σ}}_i}\exp [-\frac{1}{2}{(x-{\mathrm{\μ}}_i)}^2{\left({\mathrm{\σ}}_i^2\right)}^{-1}]---\left(2\right)$

The parameter of whole hybrid density is θ=(a ₁..., a _Mθ ₁..., θ _M).

Setting up the GMM model specifically realizes through following a, b, c step:

If power spectrum be F (i) (i=1...N), wherein N is the location number, M is the Gaussian function branches of GMM model, M get usually N/n (n=1,2,3...), tale is N _Total

A. power spectrum F (i) (i=1...N) being done normalization by following formula handles:

$f\left(i\right)=F\left(i\right)/\underset{i=1}{\overset{N}{\mathrm{\Σ}}}F\left(i\right),(i=1\·\·\·N)---\left(3\right)$

B. the signal f (i) after the normalization (i=1...N) is represented with GMM, and obtains approximate signal f ' (i) (i=1...N):

$f^{\′}\left(i\right)=s\underset{j=1}{\overset{M}{\mathrm{\Σ}}}f\left(\mathrm{sj}\right)p_j\left(i\right)=\underset{j=1}{\overset{M}{\mathrm{\Σ}}}\mathrm{sf}\left(\mathrm{sj}\right)p_j\left(i\right),(i=1\·\·\·N)---\left(4\right)$

S=N/M wherein.

$p_j\left(i\right)=\frac{1}{\sqrt{2\mathrm{\π}}\mathrm{\σ}}\exp [-\frac{1}{2{\mathrm{\σ}}^2}{(i-\mathrm{sj})}^2],(i=1\·\·\·N,j=1\·\·\·M)---\left(5\right)$

σ gets σ=1～s usually;

In fact, can calculate:

$\underset{j=1}{\overset{M}{\mathrm{\Σ}}}\mathrm{sf}\left(\mathrm{sj}\right)=1---\left(6\right)$

(6) formula satisfies the weights condition of GMM model in (1) formula:

a _i>=0, (i=1 ..., M).

C. be calculated as follows, and round the original power spectrum of back recovery:

$F\left(i\right)=f^{\′}\left(i\right)\·N_{\mathrm{total}}/\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{f}^{\′}\left(i\right),(i=1\·\·\·N)---\left(7\right)$

Like this, just accomplished the foundation of power spectrum GMM model, and with the weights of each Gaussian function in the GMM model data after as dimensionality reduction;

The gamma-spectrometric data of D after with dimensionality reduction carries out WAVELET PACKET DECOMPOSITION, and the energy of each sub-band signal carried out normalization handle, and extracts normalization data as proper vector, specifically through following a, b, the realization of c step:

The gamma-spectrometric data of a after with dimensionality reduction carries out N layer WAVELET PACKET DECOMPOSITION, and then the N layer forms 2 of equiband ^NIndividual frequency band, the signal decomposition coefficient X of extraction each frequency band from the low frequency to the high frequency _j, j=1,2 ..., 2 ^N

B is by WAVELET PACKET DECOMPOSITION coefficient X _jEach frequency band letter f of reconstruct _j(t), then resultant signal f (t) can be expressed as

$f\left(t\right)=\underset{j=1}{\overset{2^N}{\mathrm{\Σ}}}{f}_j\left(t\right)---\left(8\right)$

C asks for the energy of each frequency band, and the constitutive characteristic vector.The energy that makes j frequency band is E _j, x _JkBe reconstruction signal f _j(t) the outer discrete point amplitude of k, then

$E_j=\underset{k=1}{\overset{n}{\mathrm{\Σ}}}{\left|x_{\mathrm{jk}}\right|}^2,(j=\mathrm{1,2},...,2^N)---\left(9\right)$

N is each frequency band reconstruction signal f in the formula _j(t) discrete point number; By E _jThe constitutive characteristic vector is X=[E ₁, E ₂..., E _2N], become after the normalization

$X=[\frac{E_1}{E},...,\frac{E_{2^N}}{E}]---\left(10\right)$

Wherein $E={\left(\underset{j=1}{\overset{2^N}{\mathrm{\Σ}}}{\left|E_j\right|}^2\right)}^{1/2};$

E as sample, adopts expectation maximization method (Expectation Maximization is abbreviated as EM) that these samples are carried out interative computation proper vector, obtains the GMM model of all kinds of proper vectors, specifically realizes through following a, b step:

A tentatively sets up the GMM model of all kinds of power spectrum feature vector, X:

$P(x,\mathrm{\θ})=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{a}_i{p}_i(x;{\mathrm{\θ}}_i)---\left(11\right)$

Wherein M is the mixed number of Gaussian Mixture Model Probability Density, and x is that dimension is the power spectrum proper vector of d, p _iBe basic density, its average is μ _i, variance matrix is a ∑ _i, a _iBe to mix flexible strategy.Each basic density is the Gaussian function of d dimension, θ _iBe unknown parameter μ _iAnd ∑ _iVector representation, i.e. θ _i=(μ _i, ∑ _i), density function p _i(x, θ _i) form following:

$p_i(x,{\mathrm{\θ}}_i)=\frac{1}{{\left(2\mathrm{\π}\right)}^{d/2}{\left|{\mathrm{\Σ}}_i\right|}^{1/2}}\exp [-\frac{1}{2}{(x-{\mathrm{\μ}}_i)}^T{\left({\mathrm{\Σ}}_i\right)}^{-1}\left({x-\mathrm{\μ}}_i\right)]---\left(12\right)$

The parameter estimation of GMM adopts the b EM algorithm in step;

B adopts the expectation maximization method; Be Expectation Maximization (being abbreviated as EM); With carrying out interative computation up to convergence in all kinds of proper vector sample substitution formulas (13)～(15), realize that the renewal of gauss hybrid models parameter also finally obtains the GMM model of all kinds of samples;

$a_l^{\mathrm{new}}=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}p\left(l\right|x_i,{\mathrm{\θ}}^g)---\left(13\right)$

${\mathrm{\μ}}_l^{\mathrm{new}}=\frac{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{x}_{i}p\left(l\right|x_i,{\mathrm{\θ}}^g)}{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}p\left(l\right|x_i,{\mathrm{\θ}}^g)}---\left(14\right)$

${\mathrm{\σ}}_l^{2\left(\mathrm{new}\right)}=\frac{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}p\left(l\right|x_i,{\mathrm{\θ}}^g){(x_i-{\mathrm{\μ}}_l^{\mathrm{new}})}^2}{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}p\left(l\right|x_i,{\mathrm{\θ}}^g)}---\left(15\right)$

N is the number of all kinds of proper vector samples in formula (13)～(15), x _iBe proper vector sample, a _l, μ _lAnd σ _l ²Be respectively weight, average and the variance of each gauss of distribution function;

Go on foot the training process of promptly accomplishing power spectrum through above A～E step;

2. with the power spectrum to be identified classification of making a strategic decision, by following A～E step:

A adopts wavelet method or polynomial expression method that power spectrum to be identified is carried out smothing filtering,

B asks for the GMM model (Gaussian mixture model) of power spectrum to be identified, and with the weights of each Gaussian function in the GMM model data after as dimensionality reduction, and method is with the C step in the 1. step;

The to be identified gamma-spectrometric data of C after with dimensionality reduction carries out WAVELET PACKET DECOMPOSITION, and the energy of each sub-band signal carried out normalization handle, and extracts normalization data as proper vector, and method is with the D step in the 1. step;

D as the multidimensional random number, and calculates the class conditional probability density that it belongs to all kinds of GMM models with the proper vector of power spectrum to be identified, and method is following:

Choose the proper vector x={x of power spectrum to be identified ₁, x ₂... x _d, x ₁, x ₂..., x _dBe d kind characteristic quantity, its type conditional probability density does

$p\left(x\right|C_i)={\mathrm{\Π}}_{n=1}^{d}p\left(x_n\right|C_i),i=1,...,m---\left(16\right)$

P (x wherein _n| C _i) be C _iThe class conditional probability density function of n characteristic of class, m is the classification number;

E presses the Bayesian decision classification at last, the following a of judging process, b step:

A at first calculates $\underset{n=1}{\overset{d}{\mathrm{\Σ}}}\mathrm{Log}\left(p\left(x_n\right|C_i)\right),i=1,...,m.$

If b

Then judge x{x ₁, x ₂..., x _n∈ C _j

Go on foot the identifying of promptly accomplishing power spectrum to be identified through above A～E step.

Can find out that from above-mentioned radioactive energy spectrum recognition methods the present invention has combined the statistical property and the non-stationary property of power spectrum in the radioactivity survey, adopt wavelet method or polynomial expression method that power spectrum is carried out smothing filtering earlier in the training stage, to eliminate undesired signal; Then power spectrum is done dimension-reduction treatment, still have most information of original ability spectrum signal with the statistical property that guarantees power spectrum at lower dimensional space with the GMM model; Then, extract the normalization proper vector and, guaranteed to adopt low dimension sample to represent the uniqueness of ability spectrum signature, also guaranteed the independence of sample and Measuring Time simultaneously as sample at the multiband signal space; At last, adopt the expectation maximization method that these samples are carried out interative computation, obtain the GMM model of all kinds of proper vectors, so not only guaranteed that accurate convergence but also Gaussian function number that can be through the reasonable GMM of adjustment model are to satisfy the needs of discerning under the various different occasions.In addition, the present invention adopts the Bayesian decision classification when identification, guaranteed the optimum of recognition accuracy under statistical significance.

In a word, the present invention a kind ofly can quick training can avoid again restraining or the method for local convergence, the identification of power spectrum can be carried out smoothly and can guarantee higher recognition accuracy.

In the embodiment of the invention described above, recognition methods specifies to radioactive energy spectrum, but the need explanation is; The above is merely one embodiment of the present of invention; The present invention can discern the power spectrum of local spectral coverage equally, can be used for the identification of various ray energy spectrums, and is all within spirit of the present invention and principle; Any modification of being done, be equal to replacement, improvement etc., all should be included within protection scope of the present invention.

Claims (8)

1. radioactive energy spectrum recognition methods is characterized in that concrete steps are following:

1. the power spectrum that has recorded in the radioactivity survey is trained;

2. with the power spectrum to be identified classification of making a strategic decision.

2. a kind of radioactive energy spectrum recognition methods according to claim 1 is characterized in that, saidly the power spectrum that has recorded in the radioactivity survey is trained in 1., comprises following steps:

A carries out filtering, dimensionality reduction and classification processing to the power spectrum that has recorded in the radioactivity survey,

B extracts the proper vector of handling the back power spectrum,

C as sample, and obtains the GMM model of all kinds of proper vectors with proper vector through training.

3. a kind of radioactive energy spectrum recognition methods according to claim 2 is characterized in that, among the said A: filtering is meant adopts wavelet method or polynomial expression method that power spectrum is carried out smothing filtering; Dimensionality reduction is meant the GMM model (Gaussian mixture model) of asking for power spectrum, and with the weights of each Gaussian function in the GMM model data after as dimensionality reduction; Classification is meant by radioactive source classifies.

4. a kind of radioactive energy spectrum recognition methods according to claim 2; It is characterized in that; Extract proper vector among the said B; Be meant the data behind claim 3 dimensionality reduction are carried out WAVELET PACKET DECOMPOSITION, and the energy of each sub-band signal is carried out the normalization processing, extract normalization data as proper vector.

5. a kind of radioactive energy spectrum recognition methods according to claim 2; It is characterized in that; Obtain the GMM model of all kinds of proper vectors among the said C through training, be meant the proper vector that claim 4 is extracted, adopt expectation maximization method (Expectation Maximization as the multidimensional random number; Be abbreviated as EM) these random numbers are carried out interative computation, obtain the GMM model of all kinds of proper vectors.

6. a kind of radioactive energy spectrum recognition methods according to claim 1 is characterized in that, said 2. in the power spectrum to be identified classification of making a strategic decision, comprise following steps:

A adopts wavelet method or polynomial expression method that power spectrum to be identified is carried out smothing filtering, and asks for the proper vector of power spectrum to be identified,

B asks for the class conditional probability density of proper vector by the GMM model,

C presses the Bayesian decision classification then.

7. a kind of radioactive energy spectrum recognition methods according to claim 6; It is characterized in that; Ask for the proper vector of power spectrum to be identified among the said a; Be meant power spectrum to be identified is carried out WAVELET PACKET DECOMPOSITION after by claim 3 dimensionality reduction, and the energy of each sub-band signal is carried out normalization handle, extract normalization data as proper vector.

8. a kind of radioactive energy spectrum recognition methods according to claim 6; It is characterized in that; Ask for the class conditional probability density of proper vector among the said b by the GMM model; Be meant the power spectrum proper vector to be identified that claim 7 is extracted as multidimensional random number to be classified, and calculate the class conditional probability density that it belongs to all kinds of GMM models.

Images